ESDL: Entity Summarization with Deep Learning
Abdulkarim Kushk, Krzysztof Kochut · 2021
Entity summarization is the task of generating a general-purpose abridged description of Knowledge Graph entities. It became a valuable task and attracted a lot of attention due to the explosion in the size of semantic datasets. While most existing methods are unsupervised and do not use deep learning, in this short research paper we present ESDL, a simple yet effective supervised deep learning model. In this model, we exploit textual semantics for encoding RDF triples. As a supervised model, it uses frequency scores programmatically extracted from human-generated summaries as labels for training. Experimental results on a public benchmark dataset show that our model achieved comparable F1 scores to the state-of-the-art model DeepLENS [1] and outperformed all other models with statistical significance. Furthermore, our model has a straightforward structure and less computational complexity, resulting in ESDL comparing favorably in terms of efficiency. Experiments show that ESDL is 40% faster than DeepLENS.